Fast Facts
- Humans aren’t random: We favor 7—often because it feels less predictable—so human answers form a hill, not an even spread.
- Pretraining explains the peak, not the spike: Gradient descent learns how often people choose each number; it should copy the human-shaped hill.
- Post-training and decoding sharpen it: Preference training rewards typical answers, while low-temperature sampling makes the model’s favorite even more dominant—often 7.
- Need a genuinely random number? Use a random-number generator. An LLM is trained to give expected answers, not to roll a fair die.
Why 7 Feels Random
A viral chart asks people and an AI model to pick a number from 1 to 10. A random generator gives each number a 10% chance. People often favor 7 instead. One survey found that about one-third of respondents chose it.
That choice feels random, but it reflects familiar habits. People avoid the edges and the obvious midpoint. Meanwhile, 7 seems unusual without seeming forced. So a human distribution forms a hill around 7, not a flat line.
How a Hill Becomes a Spike
The chart gets one point right: Training helps explain why 7 leads. A model learns patterns from human language, where 7 appears as a favored answer. But gradient descent does not, by itself, explain a sharp spike. It generally teaches the model to match the patterns in its training data.
Later steps can narrow the model’s choices. Human feedback may reward answers that seem typical. Also, low-temperature settings favor the model’s most likely answer. Together, these forces can turn a gentle hill into a much sharper peak.
Still, the viral chart offers an imperfect comparison. It shows a population of people beside repeated answers from one model. And it gives no sample size or uncertainty. Treat its shape as an illustration, not a precise measurement.
Use the Right Tool
AI models can help with writing, analysis and many everyday tasks. However, they do not reliably produce random numbers. For a fair draw, use a random-number generator or a short program. Raising a model’s temperature may add variety, but it cannot guarantee fairness.
That distinction matters as more people use AI for practical work. A model can sound confident and still miss the task’s purpose. If you need a random number, use a tool built for randomness.
Continue Your Tech Journey
Learn how the Internet of Things (IoT) is transforming everyday life.
Discover archived knowledge and digital history on the Internet Archive.
AITechV1
